人脸识别赵立宏.txt

一篇人脸识别的博士论文

分类号密级

UDC 


学位论
文


人脸检测和识别算法的研究与实
现




东北大学信息科学与工程
申请学位级别
： 
博士学科类别：工学
学科专业名称
： 
模式识别与智能系统
论文提交日期：2006 年1 月10 日论文答辩日期：2006 年2 月22 日
学位授予日期：答辩委员会主席：
评阅人：

东北大学
2006 年1 月


A Dissertation for the Degree of Doctor in Pattern Recognition and 
Intelligent System 

Research and Realization in Face Detection 
and Recognition Algorithms


By Lihong Zhao 

Supervisor：Professor Xinhe Xu 

Northeastern University
January 2006



独创声
明


本人声明所呈交的学位论文是在导师的指导下完成的。论文中取得的研究成果除加
以标注和致谢的地方外，不包含其他人已经发表或撰写过的研究成果，也不包括本人为
获得其他学位而使用过的材料。与我一同工作的同志对本研究所做的任何贡献均已在论
文中作了明确的说明并表示诚挚的谢意。

学位论文作者签名：

签字日期：

学位论文版权使用授权
书


本学位论文作者和指导教师完全了解东北大学有关保留、使用学位论文的规定：即
学校有权保留并向国家有关部门或机构送交论文的复印件和磁盘，允许论文被查阅和借
阅。本人同意东北大学可以将学位论文的全部或部分内容编入有关数据库进行检索、交
流。

（如作者和导师同意网上交流，请在下方签名：否则视为不同意）

学位论文作者签名：导师签名：

签字日期：签字日期：

-I 


东北大学博士学位论文摘要

人脸检测和识别算法的研究与实现

摘
要


生物特征识别是利用人类特有的生理或行为特征来识别个人身份的技术，它提供了
一种高可靠性、高稳定性的身份鉴别途径。人脸检测和识别是目前生物特征识别中最受
人们关注的一个分支，是当前图像处理、模式识别和计算机视觉领域内的一个热门研究
课题，在公安部门罪犯搜索、安全部门动态监视识别、银行密码系统等许多领域有广泛
的应用。与指纹、视网膜、虹膜、掌纹等其他人体生物特征识别方法相比，人脸识别具
有直接、友好，使用者无心理障碍等特点。本文对此进行了较为深入的研究，论文的主
要工作和成果有以下几个方面：

⑴.. 全面概述了生物特征识别技术及其发展方向、应用背景和研究意义，重点描述
了人脸识别技术的研究内容、方法、应用前景，介绍了人脸识别技术在国内外的研究现
状，对人脸自动检测与识别技术进行了综述。
⑵.. 提出一种非线性变换的彩色空间来描述肤色模型，在该非线性彩色空间上进行
人脸肤色的分割，采用基于区域增长算法的自适应阈值处理，实现了一个完整的皮肤分
类器。通过使用自适应阈值的模糊分割技术，使皮肤区域与非皮肤区域有效地分割开，
从而得到人脸候选区域。提出利用多尺度形态边缘检测算法定位眼睛和嘴的位置，根据
均值和方差分割出的纹理特征和人脸几何特性来定位人脸，从而验证候选区域是否为人
脸。
⑶.. 在标准PCA 原理基础上，分别提出了对称主成分分析和核主成分分析算法进行
人脸识别。通过引入镜像样本，将人脸图像进行奇偶分解，并分别对奇偶图像应用KL 
展开，提取奇偶对称KL 特征；根据各个特征分量在人脸中所占能量比例的不同以及对
视角、旋转、光照等干扰的不同敏感程度，进行特征选择，增强特征的稳定性；从理论
分析入手，建立理论基础，并将该算法成功应用于人脸识别中。该算法从理论上提出奇
偶正交重构，在应用上利用镜像样本扩大样本容量，提高了识别性能并增强了人脸识别
算法的实用性。作为一类核方法，KPCA 方法在模式识别领域中得到了较多的应用，其
基础是使用KPCA 进行特征抽取。在进行非线性映射之前，首先利用经典的主分量分析
降维，然后再进行核主分量分析(KPCA)。在ORL 标准人脸库上的实验结果验证了所提
算法的有效性。
⑷.. 提出了基于小波变换图像相关性的人脸识别方法。用小波变换将原始图像分解
提取特征，可以有效地降低特征向量的维数；将训练集中的5 幅图像取平均值作为模板
脸，计算测试集中的5 幅小波变换图像与模板脸的相关系数，并进行比较。在ORL 人
脸库上的实验结果表明，提出的方法可以达到98.5% 的正确识别率，计算量小，速度快，
可用于各种人脸识别系统中。
-II 


东北大学博士学位论文摘要

⑸.. 提出融合小波特征和离散余弦变换特征的支持向量机人脸识别方法。通过小波
变换提取图像的低频分量，再利用离散余弦变换的较好压缩性能及计算的有效性提取样
本图像的特征，该方法提取的特征少而精，使输入向量的维数大大减小，减少了计算的
复杂性；同时结合支持向量机的强大分类能力，对标准的ORL 人脸库进行分类识别，
取得了很好的分类、识别效果。
⑹.. 提出将小波变换和主成分分析方法结合提取人脸特征，有效地减少了人脸图像
维数，减少了神经元网络的训练和识别时间，提高了效率；利用隐层数和隐层单元数计
算公式，合理选择神经元网络隐层数和隐层单元数，获得较好的识别结果。
由于人脸自动识别系统相对比较复杂，涉及的内容很多，本文虽然在人脸检测与识
别方面取得了一些成果，但距离实际应用还有一定的差距，有待于在今后的工作中继续
研究改进和完善。人脸自动识别是近年来非常活跃的研究领域，新思想、新技术、新方
法和新应用层出不穷，相信在不久的将来一定会找到比较完美的解决办法，到那时候人
们就可以更加充分的享受这一技术给人们的工作和生活带来的方便。

关键词：生物特征识别；人脸检测；人脸识别；小波变换；离散余弦变换；支持向量机；
主成分分析；镜像主成分分析；核主成分分析；特征提取和选择；神经元网络

-III 


东北大学博士学位论文Abstract 

Research and Realization in Face Detection and Recognition
Algorithms


Abstract 

Biometrics is a kind of science and technology using individual physiological or 
behavioral characteristics to verify identity. It provides a highly reliable and robust approach 
to the identity recognition. Automatic face detection and recognition is one of the most 
attention branches of biometrics and it is also the one of the most active and challenging tasks 
for image processing, pattern recognition and computer vision. It is widely applied in 
commercial and law area, such as mug shots retrieval, real-time video surveillance in security 
system and cryptography in bank and so on. Face recognition has direct, friendly 
characteristic s and it is no psychological obstacle for users. This dissertation mainly studies 
the approaches to frontal face detection and recognition. The main research works and 
contributions are as the following. 

⑴The biometrics technology and its development, application, and signification is 
summarized. The research content, approach and development are emphasized. The research 
status is introduced. The technology of the face detection and recognition are summarized. 

⑵A color space based on non-linear transformation is proposed. The face skin 
segmentation is finished in this non-linear color space. The face skin classification based on 
algorithms of adaptive threshold of region growing is realized. The skin regions and non-skin 
regions are separated with the fuzzy segmentation of adaptive threshold. The eyes and mouth 
are located with the multi-scale morphological algorithms. The face is located by the texture 
and geometrical features of face, then it is tested whether the candidate region is the face or 
not. 

⑶Symmetrical Principal Component Analysis (SPCA) and Kernel Principal Component 
Analysis (KPCA) are proposed based on the Classical Principal Component Analysis (CPCA). 
The face image is decomposed to odd and even images by introducing the mirror example to 
extract the odd and even symmetrical Karhunen-Loeve features. The features are select based 
on the different proportion of feature component in the face image and the different 
sensitivities in visual angle, rotation and illumination. The odd and even orthonormal 
reconstructure in the algorithm is proposed theoretically and higher correct. Recognition rate 
is achieved for the face with SPCA. Its main idea is that CPCA is first employed to preprocess 
the original training images before the nonlinear mapping and KPCA is used to extract 
features. The experimental results on ORL face databases indicate that the proposed method is 
more effective. 

⑷The paper proposes a classification method based on wavelet transform and features 
correlation. Its main idea is that the wavelet transform is first employed to preprocess the 

-IV 


东北大学博士学位论文Abstract 

original face image and reduced the dimensions of the feature space. The mean of five face 
images in training set is taken as a template faces. The correlationcoefficient is calculated and 
compared with five images of testing set and template face. The experimental results on ORL 
face databases indicate that the proposed method is more effective and the correct recognition 
rate is 98.5%. The approach is simple and faster and retains its accuracy. It is verified that 
the proposed algorithm is effective in the different application systems of face recognition. 

⑸A method of face recognition based on wavelet transform and Discrete Cosine 
Transform (DCT) and SVM is proposed. The low frequency sub-image is transformed by 
DCT, and only a small set of coefficients is retained as the features that are inputted to SVM. 
The experiments show that the performance is satisfactory. 

⑹A method of face recognition based on wavelet transform and principal component 
analysis is proposed to extract feature and reduce the dimensional feature space. The reduced 
features are inputted into BP neural network. Applying optimum algorithm of neural network 
and algebra equation theory and hidden structure basis equation, direct and indirect computing 
methods are studied to compute quantitatively the numbers of hidden layers and the unit 
numbers per hidden layer. The experiments show that the performance is perfect. 

Automatic face recognition system is comparatively complicated and involves a large 
number of contents. Although some achievement has been gained in this dissertation, more 
research is still needed to transmit the theory into practical application. More research work is 
needed to improve and perfect the methods in the future. With long term researching and 
studying, the era of intelligent acquisition and processing and application for face recognition 
will come true. 

Key words ：Biometrics, Face detection, Face recognition, Wavelet transform, Discrete Cosine 
Transform (DCT), Support Vector Machine(SVM), Principal Component Analysis(SPCA), 
Symmetrical Principal Component Analysis(SPCA), Kernel Principal Component 
Analysis(KPCA), Feature extract and select, Neural network. 

-V 


东北大学博士学位论文第一章绪论

目
录


独创声明.......................................................................................................
I
摘要..........................................................................................................II
Abstract ...................................................................................................... IV
目录........................................................................................................ VI